Questions about EcoRouter

EcoRouter is an AI efficiency layer built around one principle: use only as much AI as a task actually needs.

Today it does two things. It analyses each request and routes it to an appropriately capable model rather than defaulting to the largest one available, and it shows you the result of that decision in an EcoReceipt.

Its longer-term architecture adds a third idea, Knowledge Sequestration: preserving useful answers so that sufficiently reliable ones might eventually be retrieved rather than generated again.

Reuse what we already know. Route efficiently when generation is necessary. Reveal what actually happened.

Routing and revealing work today. Knowledge Reuse V1 — exact and conservative lexical matching — is also live; see below.

Most conversations about AI efficiency begin after a decision has already been made: a model is going to run, so how do we make it cheaper?

EcoRouter asks earlier questions.

Does this task require the most powerful model available? And eventually: does this answer need to be generated again at all?

The future of AI efficiency isn’t only making models more efficient. It’s becoming more intelligent about when we need to generate at all.

Which is the same principle stated plainly: use only as much AI as you need.

EcoRouter is built by ChatGPTree. Public attribution is “EcoRouter by ChatGPTree”.

Stumpy is ChatGPTree’s guide to using less AI and preserving more useful knowledge. Stumpy is a character, not the assistant — when you ask EcoRouter a question, EcoRouter answers.

When new generation is required, EcoRouter analyses the request and routes it to the lowest-resource capability tier expected to complete it adequately, within its routing policy.

Efficiency never means deliberately choosing a model that fails the task. An answer that doesn’t work saves nothing — it moves the cost onto the person who has to ask again.

Read the methodology for how routes, costs and comparisons are calculated.

Knowledge Sequestration is EcoRouter’s approach to preserving useful AI-generated knowledge so that, when appropriate, existing knowledge can eventually be reused rather than regenerated.

Generate once when necessary. Reuse when appropriate.

Public Knowledge is the foundation for it. When someone deliberately publishes an answer, that answer becomes a durable object other people can find.

Knowledge Reuse V1 is enabled. When a sufficiently strong existing Public Knowledge match exists — found through exact or conservative lexical matching, not yet semantic matching — EcoRouter can recommend it before generating a new answer, and you can reuse it or generate a new one anyway. Publishing continues to build the corpus this draws from.

Because they are computationally different operations. Generating a response runs accelerator-based inference and produces tokens sequentially. Retrieving an existing knowledge object can potentially use much lighter database, cache and search infrastructure.

Retrieval is not free. Databases, storage, networking, indexes, embeddings, vector search and replication all consume real resources, and a retrieval system that is never used still costs something to run.

The thesis is narrower than “retrieval is cheap”: when a sufficiently reliable existing answer can safely substitute for a new inference, retrieving it should generally require substantially less computation than generating it again.

Because the value of a preserved answer compounds with reuse.

A hundred similar requests can mean a hundred separate generations. Or one generation and ninety-nine retrievals.

Using the illustrative figures above — the published median for the generation side, a hypothetical assumption for the retrieval side — a hundred generations would be roughly 31 Wh, while one generation plus ninety-nine retrievals would be roughly 0.59 Wh.

Illustrative scenario, not measured EcoRouter savings. The conclusion worth keeping is conceptual: the computational value of useful preserved knowledge compounds as it is reused. The most efficient AI response may be the one nobody needs to generate.

Not quite, and the difference is the point.

Conventional retrieval-augmented generation usually looks like: retrieve, add context, then generate. The retrieval improves the answer — but a generation still happens.

Knowledge Sequestration aims to make another path possible: retrieve, and if the match is strong enough and the quality, freshness and safety thresholds allow, return the preserved knowledge. Otherwise generate, and potentially preserve.

The computational distinction is simply that retrieval alone does not avoid generation in a conventional RAG pipeline.

Because the goal is not maximum reuse. It is maximum appropriate reuse, without sacrificing whether the answer is any good.

A reuse decision has to weigh semantic similarity, freshness, provenance, how volatile the underlying facts are, the user’s context, how personal the request is, and confidence in the match. When those cannot establish a sufficiently strong match, the right action is to generate.

V1 of that decision system is live today, using exact and conservative lexical matching — favoring precision over recall. The fuller description above (semantic similarity, provenance, personalization) remains the longer-term direction and is not yet built.

Three categories, kept separate on purpose:

Measured — directly observed from the provider interaction, such as tokens and latency. Calculated — arithmetic over measured values and configured reference data, such as cost. Estimated — requires assumptions, such as estimated relative compute reduction.

Read the methodology →

No.

Providers do not publish per-request energy draw, which hardware a request landed on, how utilised it was, or the carbon intensity of the grid supplying it at that moment. Without those, a number in watt-hours or grams of CO₂ would be a figure with a unit attached rather than a measurement.

So EcoRouter reports a comparative estimate — estimated relative compute reduction against a configured frontier baseline — and says that it is an estimate. We would rather be transparently approximate than precisely wrong.

We don’t make either claim. Substantiating them would require lifecycle accounting we do not have.

EcoRouter’s environmental methodology is about computational efficiency: routing to a smaller adequate model, and estimating the relative difference that makes. Separately, ChatGPTree supports ecological restoration as part of its mission. Those are two different things, and we do not attribute a particular tree to a particular inference.

Not by itself.

Publishing preserves an answer so it can potentially become reusable knowledge. A computational benefit arises only when future generation is actually avoided — which can now happen: an eligible public entry can be recommended instead of generating a new answer when a sufficiently strong match exists.

Publishing creates the opportunity. It does not bank the saving.

Knowledge Sequestration is an engineering thesis, and we intend to test it rather than assert it. The things that would need measuring include the actual resource cost of our retrieval path, how often a preserved answer can satisfy a new request, how confident the match is, how often the system has to escalate to generation anyway, and whether reused answers are as good.

One metric we may eventually report is an inference avoidance rate: the share of eligible requests satisfactorily fulfilled from preserved knowledge without new generative inference.

That is a proposed future metric. EcoRouter does not report it today, and there is nothing yet to report.

A library of question-and-answer pairs that people deliberately chose to preserve. From a private chat, nothing is published automatically: you review exactly what would become public and confirm it explicitly.

Browse Public Knowledge →

Yes. A public Channel is a shared knowledge space: questions asked there, and their answers, become Public Knowledge unless removed. Answers that used private file context are never published, and private Channels stay private to their members.

Share creates an unlisted link to the selected question and answer. Anyone holding that link can read it, but it is not listed in Public Knowledge, not in the sitemap, and not indexed.

Make public deliberately contributes that question and answer to the Public Knowledge corpus, where anyone can find it.

An unlisted link is not private. Possession of the URL grants access — that is what “unlisted” means, and it is why we don’t call it private.

Publishing operates on one selected question and its answer. Earlier and later turns in the conversation are not part of the published object.

Before publishing, EcoRouter runs a check for obviously sensitive content — things like email addresses, phone numbers and credentials — shows you exactly what would become public, and requires explicit confirmation. That check looks for recognisable patterns; it cannot understand what is sensitive to you, so the review step is the one that matters.

ChatGPTree is the company behind EcoRouter. It is exploring ways for AI to use fewer computational resources, preserve useful knowledge, and support real-world ecological restoration.

EcoRouter is the part of that work concerned with using less AI per task and being transparent about it.